AI Treaties Need Verification Tech Before Trust Can Follow

The verification challenge facing AI governance

When Robert Oppenheimer declared in 1965 that stopping nuclear proliferation was "20 years too late," he was wrong. Despite his pessimism, nuclear non-proliferation succeeded—but not through trust alone. The breakthrough came from verification technology that let adversaries confirm compliance without revealing sensitive secrets.

AI governance faces an identical challenge today, according to Tim Fist, director of emerging technology policy at the Institute for Progress, and Janet Egan, senior fellow at the Center for New American Security. The details were first reported by the Center for Humane Technology's AI Watch.

The recent Trump-Xi Summit opened a policy window for US-China coordination on AI safety, driven partly by concerns over models like DeepSeek's Mythos and its cyber-offensive capabilities. But expressing interest in coordination is easy. Actually trusting a geopolitical rival to follow through requires something concrete.

How nuclear verification worked

Two historical examples illuminate the path forward. The Comprehensive Test-Ban Treaty became possible only after 300 seismic monitoring stations were deployed globally to detect underground nuclear tests. Without that technology, no agreement could be verified.

Even more instructive: the Intermediate-Range Nuclear Forces Treaty relied on x-ray scanning technology called Cargo Scan, developed jointly by the US and Soviet Union. Placed at Soviet missile factories, these scanners measured the diameter of missiles in rail cars—enough to verify compliance, but not enough to reveal other design secrets. The technology was deliberately privacy-preserving, allowing both parties to share information without compromising national security.

The compute bottleneck as verification infrastructure

For AI, the most promising verification target is compute—the physical chips required to train advanced models. Unlike algorithms or training data, which can be copied instantly, chips represent a tangible bottleneck with a narrow, largely US-controlled supply chain.

Modern chips from Intel, AMD, and NVIDIA already ship with "trusted execution environments"—secure vaults within the hardware that can create cryptographic fingerprints of what programs actually run. When a model trains inside this vault, the hardware generates a signed statement proving exactly what computation occurred. This capability exists today, though its application to treaty verification remains nascent.

Why it matters

Leading AI labs—DeepMind, Anthropic, and OpenAI—have all stated they would support a global slowdown in AI development if recursive self-improvement becomes imminent. But no lab will slow down unilaterally while competitors race ahead. The window for US-China coordination is open now, yet without verification infrastructure in place, any agreement remains theoretical. Building that infrastructure takes years. The nuclear precedent shows that waiting for consensus on what to govern before developing how to verify it wastes critical time. The technology must be ready when political will materializes—not developed afterward.

The path forward

AI's global impacts will cross borders regardless of which country develops transformative capabilities first. Risks that emerge in one jurisdiction don't respect national boundaries. In an era of diminishing trust between international counterparts, verification technology becomes the foundation for any meaningful coordination.

The nuclear example offers both hope and urgency: verification technology can enable agreements that seem politically impossible. But it must be built before the critical moment arrives, not after.

This analysis draws from a conversation between Tristan Harris and experts Tim Fist and Janet Egan, originally published by the Center for Humane Technology.